A full walk of the registry: 22,636 active, 12,077 with a reachable endpoint (data + tool)

[ogasurfproject-jpg]

Pre-submission Checklist

• I have verified that this discussion would not be more appropriate as an issue in a specific repository
• I have searched existing discussions to avoid duplicates

Discussion Topic

I walked the whole registry on 2026-08-19 to answer a question I could not
find an answer to: how many registered servers actually declare an address
you can reach from outside.

229 pages, limit=100&version=latest, stopping because the cursor ran out.
Read only, 0.15s between pages, one client. None of the servers themselves
were contacted.

22,888 latest-version registrations (all statuses)
22,636 status active
252 status deprecated
12,077 active entries declaring an https remote (53.4%)
12,429 distinct endpoint URLs
9,109 distinct hosts
15,277 active entries published in the last 90 days
7,738 published in the last 30 days

Raw JSON, the full endpoint list, the counting script and the sha256:
https://github.com/ogasurfproject-jpg/mcp-registry-survey

Three observations, offered as observations rather than as complaints.

1. Concentration. 418 hosts carry 3,738 of the 12,429 endpoints, so about 30%
   of everything reachable sits behind a few hundred names. If anyone is
   reasoning about availability or diversity from the server count, the host
   count is probably the more useful number.

2. Growth. Two thirds of the active entries are younger than 90 days. Any
   analysis of this registry has a very short shelf life.

3. Status. 252 deprecated out of 22,888. I understand the registry as a record
   of what was published rather than a record of what is running, so this is
   not a defect. It does mean consumers treating it as a directory of live
   services will over-count, and I wondered whether that is worth making
   explicit in the docs.

Two questions.

First, is version=latest the intended way to get one row per server? Without
it the cursor is name:version and a naive walk counts versions, which is how I
got a wrong number on my first three attempts.

Second, I am planning a liveness pass: initialize plus tools/list against those
12,429 endpoints, read only, no tool calls. What rate would you consider
acceptable? I would rather be told here than find out from someone's error
budget.

Happy to correct anything above. If someone gets a different count I will
publish both.